Vertex AI search filtering

[Ian_Stack]

Problem: Currently have a Vertex AI search app connected to datastore.  This datastore is from a website and collects all the 'support' documents. However the response given is not always accurate.

How do I filter my search results to the ideal results? Specifically for the most seached words (refund, cancel, support, customer service.) Ideally, when one of the most searched words is entered, the summary and provided results are specific and I am able to determine what is returned

I see possible solutions:
Search Tunning and Datastore Schema with meta tags

I have tried both have questions as well.  For the search tuning, for the corpus file, query, and training file, I see that it should have at least 10,000 segments. Currently I am just copying content from the url connected to the database and just building out segments this way.  Is the process that is expeted?

For Datastore schema,  I see that the articles already have some meta tags on them.  However I still. don't understand how to reference these meta tags in the schema and then configure them after.  I am using the widget, and I don't know where to add a 'booster' or any configuration in the console to use the schema.

would appreciate advice on both paths!

Thank you!

[dawnberdan]

Hi @Ian_Stack,

Welcome to Google Cloud Community!

Your proposed solutions are a great starting point. Let's break down how to improve your Vertex AI search results by focusing on those high-frequency keywords.

1. Search Tuning

• Corpus File: You're correct that a corpus file with at least 10,000 segments is generally recommended for good search tuning. However, simply copying content from URLs might not be the most effective approach.

• Preprocessing: Before segmenting your data, consider preprocessing it to improve search performance. This includes:

• Removing Stop Words: Eliminate common words like "the," "a," and "an" that don't add significant meaning.
• Stemming/Lemmatization: Convert words to their base form (e.g., "running" becomes "run")
• Tokenization: Split text into individual words or phrases for better analysis.

• Segment Creation: Rather than simply copying content, create segments that represent meaningful units of information, such as a paragraph or a section of a support document. 

• Query File: Your query file should contain a representative set of queries that users are likely to make. Include variations of your high-frequency keywords (e.g., "refund," "cancel," "support," "customer service").

2. Datastore Schema with Meta Tags:

• Schema Definition: Define a schema for your datastore that incorporates meta tags to specify the relevance of certain fields for search. For example, if your support documents include a "category" meta tag, you can define it in the schema and assign it a higher "boost" value. This increases the likelihood that documents with the "refund" category will appear in search results when the query includes "refund."

3. Combining Search Tuning and Schema:

• Boosting: Use the schema to boost documents that contain your high-frequency keywords. For example, you can assign a higher boost to documents that have the "refund" category or that contain the word "refund" in the title or body.
• Query Expansion: Use search tuning to expand your queries with synonyms and related terms. This can help improve the relevance of search results, especially when users use less specific terms.

For more information, you can check these resources:

• Vertex AI
• Vertex AI Search
• Data Stores
  
  

In addition, I came across an article/blog that explores the fine-tuning of LLMs. It explores the necessity, methods, and outcomes of fine-tuning language models. This might be helpful to you.

I hope the above information is helpful.